Sitecomb, a hosted MCP server for automated website quality audits, has detailed the technical and regulatory hurdles of deploying a single backend to both Anthropic's Claude and OpenAI's ChatGPT. The service, which evaluates sites across six dimensions including security, privacy, and mobile responsiveness, reveals that model providers are enforcing starkly different rules regarding commercial upselling. This divergence forces developers to implement conditional logic within their server definitions, turning a simple tool deployment into a compliance exercise.
Commercial Policy Friction
The core challenge lies in the conflicting guidelines between the two major AI platforms. As of September 30, 2026, OpenAIβs app rules prohibit upselling digital products or displaying prices, whereas Claudeβs directory permits linking to paid full-quality reports. To navigate this, Sitecomb employs a single server definition mounted twice with a tiny policy object. This object toggles features like reportLink based on the host, ensuring that the ChatGPT mount remains compliant by suppressing any direct or indirect mentions of paid services in the tool description itself.
Architectural Decisions for Compliance
Beyond policy, the engineering choices reflect the constraints of a stateless, serverless environment. Running on Next.js with mcp-handler and Streamable HTTP, Sitecomb creates a fresh server instance for every request. This approach avoids session storage complexities but requires careful management of function limits; the team set maxDuration to 120 seconds to accommodate the heavy lifting of loading a site like a real browser. The architecture prioritizes cheap refusals, checking input validity and daily limits before triggering expensive compute resources.
Guarding Against Model Hallucination
A critical aspect of the implementation is strict instruction adherence to prevent the LLM from fabricating data. Sitecombβs server instructions explicitly command the model to report findings exactly as given, forbidding the invention of additional problems or quality scores. This is vital because a model might otherwise extrapolate issues from a low score, such as assuming large images are the cause of poor performance. If a site blocks automated checks, the tool returns no score rather than a hallucinated one, maintaining trust in the audit's accuracy.
Key Takeaways
- OpenAI and Anthropic enforce different rules on upselling, requiring conditional logic in MCP server definitions.
- Sitecomb uses a stateless Next.js architecture with strict time limits to handle heavy website audits.
- Server instructions must explicitly prevent models from hallucinating quality scores or additional site issues.
- Tool descriptions should be written for natural language queries, not just technical function names.
The Bottom Line
The MCP ecosystem is rapidly maturing, but the lack of standardized commercial policies across major platforms is creating unnecessary fragmentation for developers.